Distracted Driving:

Detecting Texting with

Neural Networks


Joseph Blubaugh

29 March 2017

Contents


  • Data Introduction, Preparation, and Project Management
  • Exploratory Analysis and Model Proposal
  • Understanding Basic Neural Nets
  • Model Training and Selection
  • Exploring Model Effects

Data Introduction

Data Extraction and Preparation

Project Management


Data Introduction

  • The data are driving simulations for 66 individuals ranging from 3,000 to 30,000 observations per simulation
  • Each observation contains likelihood scores for 8 facial expressions recorded at fixed intervals of .03 seconds.
  • Stimuli data which records events introduced into each simulation and basic demographic data are also available.
  • There are over 6.7 million observations in the entire dataset spread accross 777 files.

T001-001.xlsx (Subject 01, Simulation 01)
Drawing


Data Extraction

  • Python scripts were used to extract and combine the 509 driving simulation files and 267 stimuli files into combined data sets.
  • The subject-simulation identifier was the name of each file. A column labeled ID was created based on the file name to identify the original data set.

Sample of data-faces.csv

Frame Time Anger Contempt Disgust Fear Joy Sad Surprise Neutral ID
0 0.0000 0.0101 0.0218 0.0043 0.0541 0.5260 0.0959 0.0010 0.2868 T001-001
1 0.0333 0.0101 0.0218 0.0043 0.0541 0.5260 0.0959 0.0010 0.2868 T001-001
2 0.0667 0.0101 0.0218 0.0043 0.0541 0.5260 0.0959 0.0010 0.2868 T001-001
3 0.1000 0.0080 0.0187 0.0032 0.0375 0.5353 0.1050 0.0011 0.2911 T001-001
4 0.1333 0.0091 0.0380 0.0158 0.0036 0.6902 0.0177 0.0004 0.2252 T001-001
5 0.1667 0.0104 0.0450 0.0139 0.0030 0.7157 0.0162 0.0003 0.1955 T001-001

Sample of data-stimuli.csv

Start End Event.Switch Event.Type Event ID
86.5 246.50 1 1 Analytical Questions T001-005
508.5 657.50 1 2 Mathematical Questions T001-005
107.5 269.25 1 3 Emotional Questions T001-006
521.0 674.75 1 3 Emotional Questions T001-006
81.0 240.00 1 4 Texting T001-007
510.0 671.00 1 4 Texting T001-007

  • NOTE: 2 simulation files started on different rows than the rest of the 507 files and had to be manually corrected.
    • T034-005.xlsx: header starts on row 8
    • T009-006.xlsx: header starts on row 10

Data Preparation

  • Stimuli data captured the starting and ending times of events.
  • A loop function was written to go through each record and compare time between the simulation and the starting/ending time of the event.
  • If the simulation time was between the starting and ending time of an event, then all of the records in the time interval were coded with that event (ie: Texting)
  • All observations outside of the event time interval were coded as No Event.


Sample of Cleaned Data Showing an Event Transition

Subject Trial Age Gender Frame Time Event.Switch Event Action Anger Contempt Disgust Fear Joy Sad Surprise Neutral
T001 007 Y M 2427 80.900 0 No Event 0 0.0909 0.0575 0.4205 3e-04 0.0011 0.1343 0 0.2954
T001 007 Y M 2428 80.933 0 No Event 0 0.0612 0.0397 0.4293 4e-04 0.0011 0.1630 0 0.3052
T001 007 Y M 2429 80.967 0 No Event 0 0.1034 0.0963 0.3186 2e-04 0.0013 0.0856 0 0.3946
T001 007 Y M 2430 81.000 1 Texting 4 0.0363 0.4976 0.0171 1e-04 0.0024 0.0069 0 0.4396
T001 007 Y M 2431 81.033 1 Texting 4 0.0059 0.7285 0.0027 4e-04 0.0068 0.0063 0 0.2493
T001 007 Y M 2432 81.067 1 Texting 4 0.0058 0.6890 0.0035 4e-04 0.0077 0.0068 0 0.2868


  • NOTE: The average texting event lasted 2.5 minutes, but we dont really know what occurred during the event time interval. Was there one long texting action or was the event made up of a series of sending and receiving texts?

Project Management


Reproducible Research

  • Code, plots, and this presentation are organized and hosted in a github repository.
  • The main page includes steps to reproduce the data set and models.
  • The entire dataset is too large to be hosted on github and would need to be retrieved elsewhere.
  • The texting dataset needed to reproduce this analysis is available on the git repository.


Github Project: https://github.com/jestonblu/driving

Github Project Page Drawing



Exploratory Analysis

and

Model Proposal


Exploratory Analysis

Drawing

  • The yellow and gray points represent events during the trials
  • The baseline trial has no events and is gray throughout
  • LOESS (Local Polynomial Regression) lines display the moving average over the entire simulation
  • Many subjects displayed visual differences between the texting simulation and the baseline simulation

Exploratory Analysis

Drawing

Drawing

Drawing

Drawing

Exploratory Analysis

Drawing

  • The same LOESS lines were used to show all 59 subjects on a single plot
  • All observations were centered on the overall average of the baseline simulation
  • Anger, Contempt, Disgust, and Neutral displayed more variation than the other emotions for both trials

Model Proposal


Drawing


Takeaways

  • Differences in variation between the trials suggest that it may be possible to build a model capable of predicting a texting event
  • Subject specific plots are unique enough that a variable for subject may be needed in modeling

Baseline Trial

  • Trial 4 was used as a baseline trial because the conditions were identical to the Texting Trial (dense traffic with detour).
  • The overall mean for each Subject's emotion in the baseline trial was subtracted from every observation in the Texting Trial.

Model Proposal

  • Train a Feed Forward Neural Network using emotional likelihoods and demographics to predict when a subject is texting

Neural Network Advantages

  • NNets are well suited for large data sets of continuous variables
  • Analogous to logistic regression and appropriate for predicting probabilities



Understanding Basic

Neural Networks


Neural Network Basics

Basic Neural Network Example Drawing

General Model Form

\[ nnet(O1 \sim X1 + X2, size = 3) \]


Feed-Forward Neural Networks

  • Class of Statistical Learning model
  • Uses a training set for tuning the model and a testing set for measuring performance
  • Similar to logistic regression
  • Typically displayed as a diagram of connected nodes


Neural Network Components

  • Nodes:
    • Input Nodes: Input values of the predictor variables
    • Hidden and Output Nodes: Value are the sumproduct of the connected weights
  • Weights: Represents the transformation that takes place between nodes
  • Activation Function: Transforms the output into an appropriate scale
    • For logistic regression, the sigmoid function: \( S(x) = \frac{1}{1 + \exp(-x)} \)

Neural Network Basics

Step 1: Model is Initialized with Random Weights


Step 2: Calculate Hidden Weights and Output Node Prediction

  • Hidden Node values are the sum product of the connected weights and input nodes

\[ \begin{align} H1 = & (1)(.2) + (1)(.4) = 0.6, \text{ } S(0.6) = .645 \\ H2 = & (1)(.1) + (1)(.6) = 0.7, \text{ }S(0.7) = .668 \\ H3 = & (1)(.7) + (1)(.3) = 1.0, \text{ }S(1.0) = .731 \\ \end{align} \]

  • Output Node Prediction

\[ O1 = (.645)(.3) + (.668)(.5) + (.731)(.7) = 1.039 \\ S(1.039) = .739 \]

  • Model Error: .739

NOTE: Activation Function
\[ \begin{align} S(x) = & \frac{1}{1 + \exp(-x)} \end{align} \]

First Iteration of a Basic Neural Network Drawing

NOTE: Grayed values did not change from previous step

Neural Network Basics

Step 3: Update Weights Based on Error

  • Update Weights between Hidden Layer and Output Node

\[ \begin{align} Delta = & S'(.739) = .2187 \\ \\ DeltaChange = & (.2187) / [.645, .668, .731] = [.339, .327, .299] \\ \\ w_7 = & .645 - .339 = .306 \\ w_8 = & .668 - .327 = .341 \\ w_9 = & .731 - .299 = .432 \\ \end{align} \]

  • Update Weights between Input Node and Hidden Layer

\[ \begin{align} DeltaWeights = & Delta / [.3, .5, .7] * S'([.6, .7, 1]) \\ = & [.167, .097, .061] \\ \\ DeltaChange = & DeltaWeights / [1, 1] \\ = & [.167, .097, .061, .167, .097, .061] \\ \\ w_1 = & .2 - .167 = .033 \\ w_2 = & .1 - .097 = .003 \\ w_3 = & .7 - .061 = .639 \\ w_4 = & .4 - .167 = .233 \\ w_5 = & .6 - .097 = .503 \\ w_6 = & .3 - .061 = .239 \\ \end{align} \]

Step 4: Repeat step 2 to update node values

  • S(.633) = .653, Error = .653 vs previous .739

First Iteration of a Basic Neural Network Drawing

  • Grayed values did not change from previous step
  • \( S'(x) = S(x)(1 - S(x)) \)



Model Fitting and Selection


Model Fitting

Neural Network Model Design Drawing



General Model Form

\[ \begin{align*} nnet(Texting \sim & \text{ } Subject + Age + Gender + Anger + Contempt \text{ } + \\ & \text{ } Digust + Fear + Joy + Sad + Surprise + Neutral)\\ \end{align*} \]

Modeling Strategy

  • Train the same general model on various slices of the data to see what works best

  • 12 training and testing sets were created from the combination of Data Processing and Data Split methods

  • Data Processing

    • Original: Emotions in the original form measured in .03 second intervals.
    • Differencing: First order differening of the original observations.
    • Moving Avg: Moving averages n=30 for all of the emotions.
    • ½ Sec Cut: Time cut into ½ second intervals with the average value recorded.
    • ½ Sec Diff: First order differncing of the ½ second cut data.
    • ½ Sec Cut Stat: ½ Sec Cut with additional sd, min, max, iqr, and median.

  • Data Split

    • 365 Split: The data are split at the 365 second, approximately half way through the texting simulation.
    • Entire Sim: The training set is randomly selected from the entire simulation.

Model Fitting

Statistical Software

  • R's nnet package for feed-forward neural networks

  • The Caret Package

    • Caret is a modeling framework for training classification and regression models
    • Uses models from other packages and offers a rich set of validation test and diagnostic plots
    • Can implement parallel processing of cross validation tasks

Performance and Validation Testing

  • k=10 cross validation for training sets
  • AUC (Area Under Curve) and total accuracy

Model Search Parameters

  • Max Iterations: The number of iterations allowed for training
    • 100 (250,500 and 1000 iterations are run for the best models)
  • Size: The number of nodes in the hidden layer. Increases training time exponentially
    • [1, 10, 25, 50]
  • Decay: A penality applied to weights after each iteration. Moves weights that dont update towards zero.
    • [0, .1, .2]

NOTE: Each Model is trained 120 times

  • (k=10 cross validation) x (12 combinations of size and decay)

Model Selection


Model Performance with 100 Iteration Limit

Model Data Processing Data Split MaxItr Size Decay Training Testing AUC
Model 1: Original 365 Split 100 50 .20 .760 .676 .734
Model 2: Original Entire Sim 100 50 .20 .754 .754 .847
Model 3: Differencing 365 Split 100 10 .00 .518 .516 .526
Model 4: Differencing Entire Sim 100 25 .10 .572 .571 .637
Model 5: Moving Avg 365 Split 100 10 .00 .503 .502 .527
Model 6: Moving Avg Entire Sim 100 10 .00 .528 .528 .544
Model 7: ½ Sec Cut 365 Split 100 50 .10 .820 .698 .761
Model 8: ½ Sec Cut Entire Sim 100 50 .20 .788 .779 .868
Model 9: ½ Sec Diff 365 Split 100 50 .10 .633 .602 .650
Model 10: ½ Sec Diff Entire Sim 100 50 .20 .682 .622 .681
Model 11: ½ Sec Cut Stat 365 Split 100 50 .10 .846 .716 .781
Model 12: ½ Sec Cut Stat Entire Sim 100 50 .20 .820 .803 .891

Additional Training for Best Models

Model Data Processing Data Split MaxItr Size Decay Training Testing AUC
Model 8: ½ Sec Cut Entire Sim 250 50 .10 .816 .804 .893
Model 8: ½ Sec Cut Entire Sim 500 50 .10 .828 .810 .899
Model 8: ½ Sec Cut Entire Sim 1000 50 .10 .842 .820 .906
Model 12: ½ Sec Cut Stat Entire Sim 250 50 .10 .858 .823 .906
Model 12: ½ Sec Cut Stat Entire Sim 500 50 .20 .864 .823 .907
Model 12: ½ Sec Cut Stat Entire Sim 1000 50 .10 .871 .824 .908

NOTE: Blue indicates best models



Drawing


Drawing

Model Fitting

Model Training and Validation

## Set Cross Validation
fit.control = trainControl(method = "cv", number = 10)

## Create combination of model parameters to train on
search.grid = expand.grid(decay = c(0, .1, .2), 
                          size = c(1, 10, 25, 50))

## Limit the iterations and weights each model can run
maxIt = 1000; maxWt = 15000

fit = train(Texting ~ . - Time, mdl.08.train, 
            method = "nnet",
            trControl = fit.control,
            tuneGrid = search.grid,
            MaxNWts = maxWt,
            maxit = maxIt)

44503 samples, 12 predictors, 2 classes: '0', '1' 

Resampling: Cross-Validated (10 fold) 
Summary of sample sizes: 40053, 40053, 40052, 40052, ...
Resampling results across tuning parameters:

  ------------------------------
  Decay  Size  Accuracy   Kappa
  ------------------------------
  0.0     1    0.6654     0.3042
  0.0    10    0.7857     0.5519
  0.0    25    0.8135     0.6129
  0.0    50    0.8252     0.6375
  0.1     1    0.6830     0.3182
  0.1    10    0.8052     0.5934
  0.1    25    0.8247     0.6352
  0.1    50    0.8304     0.6472 ## Best Model
  0.2     1    0.6809     0.3126
  0.2    10    0.8033     0.5889
  0.2    25    0.8196     0.6242
  0.2    50    0.8241     0.6336

Parameter Comparisons
Drawing

Confusion Matrix and Statistical Summaries

            Reference             Reference
             0     1                0    1
    Pred 0 22736  4616     Pred 0 .885 .245
         1  2943 14208          1 .114 .754  

               Accuracy : 0.8301          
                 95% CI : (0.8266, 0.8336)
    No Information Rate : 0.577           
                  Kappa : 0.6472

            Sensitivity : 0.8854          
            Specificity : 0.7548          

         Pos Pred Value : 0.8312          
         Neg Pred Value : 0.8284          
      Balanced Accuracy : 0.8201

  Area Under Curve (AUC): 0.906

Model Performance

Drawing

  • Each point is colored by the prediction of the best model
  • The blue line is a LOESS smoother of the probability predition for that corresponding prediction
  • The shaded regions represent the actual texting window
  • Yellow points within the gray regions represent correct predictions

Model Performance

Drawing

Drawing

Drawing

Drawing

Model Performance

Balanced Accuracy by Subject

T022 T007 T086 T006 T018 T083 T035 T076 T081 T064 T020 T012 T074 T088 T013 T009 T032 T003 T011 T080 Top 20
Train 0.982 0.920 0.956 0.941 0.935 0.956 0.951 0.949 0.927 0.921 0.932 0.923 0.923 0.938 0.896 0.911 0.915 0.899 0.902 0.896 .943
Test 0.975 0.951 0.950 0.941 0.936 0.930 0.928 0.927 0.922 0.921 0.919 0.911 0.907 0.900 0.898 0.897 0.892 0.884 0.874 0.872 .938
GenderMale 0 1 0 0 0 1 0 1 1 0 0 0 1 1 1 1 1 1 1 0 11
AgeOld 0 0 1 0 0 1 1 1 0 0 0 0 1 1 0 0 1 0 0 0 7
T016 T005 T060 T044 T015 T008 T079 T073 T039 T082 T010 T066 T029 T046 T024 T051 T017 T042 T001 T061 Mid 20
Train 0.893 0.865 0.909 0.886 0.864 0.868 0.893 0.855 0.869 0.869 0.849 0.849 0.813 0.857 0.810 0.856 0.807 0.843 0.851 0.793 .854
Test 0.862 0.861 0.856 0.849 0.847 0.841 0.838 0.834 0.833 0.832 0.826 0.813 0.812 0.810 0.808 0.806 0.806 0.803 0.800 0.787 .826
GenderMale 0 1 0 1 1 0 0 0 0 1 0 0 0 1 0 1 1 1 1 1 10
AgeOld 0 0 0 1 0 0 0 1 1 0 0 0 1 1 0 1 0 1 0 0 7
T084 T077 T036 T031 T033 T040 T021 T014 T019 T004 T002 T023 T054 T041 T034 T047 T025 T038 T027 Bot 19
Train 0.806 0.794 0.789 0.810 0.768 0.782 0.802 0.815 0.742 0.728 0.717 0.748 0.741 0.698 0.688 0.703 0.699 0.559 0.540 .733
Test 0.780 0.770 0.766 0.765 0.751 0.750 0.748 0.748 0.734 0.721 0.719 0.706 0.698 0.697 0.677 0.670 0.660 0.532 0.514 .705
GenderMale 1 0 1 0 0 1 1 0 1 0 0 1 1 0 1 0 0 1 0 9
AgeOld 1 1 1 1 1 1 0 0 0 0 0 0 1 1 1 1 1 1 1 13

Proportional Summary

% Male % Female % Old % Young % Young Male % Young Female % Old Male % Old Female
Top 20 36.6% 31.0% 25.9% 40.6% 40.0% 37.5% 35.7% 16.7%
Mid 20 33.3% 34.5% 25.9% 40.6% 40.0% 43.8% 21.4% 25.0%
Bot 19 30.0% 34.5% 48.1% 18.7% 20.0% 18.8% 42.9% 58.3%



















  • 14 of 59 had testing performance > 90%
  • 39 of 59 had testing performance > 80%
  • 7 of 59 had testing performance < 70%
  • The 7 worst performing Subjects were all Old

Exploring Modeling Effects

Drawing

Evaluating Differences in Age and Gender 

******************************************************************
Levene's Test for Homogeneity of Variance (Center = Median)
******************************************************************
      Df F value Pr(>F)
group  1  0.2034 0.6537
      57  

******************************************************************
General Linear Model
******************************************************************
Deviance Residuals: 
      Min         1Q     Median         3Q        Max  
-0.269925  -0.047361   0.004553   0.066577   0.184431  

Coefficients:
              Estimate Std. Error t value Pr(>|t|)    
(Intercept)   0.76528    0.02732  28.011   <2e-16 ***
Old Male      0.03652    0.03665   0.996   0.3234    
Young Female  0.08808    0.03568   2.468   0.0167 *  
Young Male    0.07124    0.03665   1.943   0.0571 .  
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

(Dispersion parameter for gaussian family taken to be 0.008957086)

    Null deviance: 0.55633  on 58  degrees of freedom
Residual deviance: 0.49264  on 55  degrees of freedom
AIC: -104.91

Number of Fisher Scoring iterations: 2

******************************************************************
Shapiro-Wilk Normality Test
******************************************************************
data:  mdl$residuals
W = 0.97066, p-value = 0.1645

Conclusions

  • Neutral, Surprise, and Anger were the most important emotions for successfully identifying texting.

  • Joy, Contempt, and Fear are the least important emotions for identifying texting.

  • Young Females had the best overall testing performance while Older Females had the worst overall testing performance.

    • At the 95% confidence level, the testing accuracy between Young Females and Old Females are different.
    • At the 90% confidence level, the testing accuracy between Young Males vs Old Females are different.

  • The variable importance plot from the training model indicated that Gender is more important than Age. The testing accuracy results by subject somewhat contradict this claim.

  • After extending the training iterations, the difference between Model 8 (½ Sec Cut) and Model 12 (½ Sec Cut Stat) are negligible. This suggests that there is much more information in the average likelihood score than in the other descriptive statistics (sd, min, max, iqr).